PFAFFIAN SCHUR PROCESSES AND LAST PASSAGE PERCOLATION IN A HALF-QUADRANT
JINHO BAIK, GUILLAUME BARRAQUAND, IVAN CORWIN, AND TOUFIC SUIDAN
Abstract. We study last passage percolation in a half-quadrant, which we analyze within the framework of Pfaffian Schur processes. For the model with exponential weights, we prove that the fluctuations of the last passage time to a point on the diagonal are either GSE Tracy-Widom distributed, GOE Tracy-Widom distributed, or Gaussian, depending on the size of weights along the diagonal. Away from the diagonal, the fluctuations of passage times follow the GUE Tracy- Widom distribution. We also obtain a two-dimensional crossover between the GUE, GOE and GSE distribution by studying the multipoint distribution of last passage times close to the diagonal when the size of the diagonal weights is simultaneously scaled close to the critical point. We expect that this crossover arises universally in KPZ growth models in half-space. Along the way, we introduce a method to deal with diverging correlation kernels of point processes where points collide in the scaling limit.
Contents 1. Introduction 1 2. Fredholm Pfaffian formulas8 3. Pfaffian Schur process 13 4. Fredholm Pfaffian formulas for k-point distributions 23 5. Convergence to the GSE Tracy-Widom distribution 33 6. Asymptotic analysis in other regimes: GUE, GOE, crossovers 46 References 56
1. Introduction 2 In this paper, we study last passage percolation in a half-quadrant of Z . We extend all known results for the case of geometric weights (see discussion of previous results below) to the case of exponential weights. We study in a unified framework the distribution of passage times on and off diagonal, and for arbitrary boundary condition. We do so by realizing the joint distribution of last passage percolation times as a marginal of Pfaffian Schur processes. This allows us to use powerful methods of Pfaffian point processes to prove limit theorems. Along the way, we discuss an issue arising when a simple point process converges to a point process where every point has multiplicity two, which we expect to have independent interest. These results also have consequences about interacting particle systems – in particular the TASEP on positive integers and the facilitated TASEP – that we discuss in [2]. Definition 1.1 (Half-space exponential weight LPP). Let wn,m be a sequence of i.i.d. n>m>0 1 exponential random variables with rate 1 when n > m + 1 and with rate α when n = m. We
1 The exponential distribution with rate α ∈ (0, +∞), denoted E(α), is the probability distribution on R>0 such −αx that if X ∼ E(α), P(X > x) = e . 1 m
w22
w11 w21 w31 n
Figure 1. LPP on the half-quadrant. One admissible path from (1, 1) to (n, m) is shown in dark gray. H(n, m) is the maximum over such paths of the sum of the weights wij along the path. define the exponential last passage percolation (LPP) time on the half-quadrant, denoted H(n, m), by the recurrence for n > m, ( n o max H(n − 1, m),H(n, m − 1) if n > m + 1, H(n, m) = wn,m + H(n, m − 1) if n = m with the boundary condition H(n, 0) = 0.
Remark 1.2. It is useful to notice that the model is equivalent to a model of last passage percola- tion on the full quadrant where the weights are symmetric with respect to the diagonal (wij = wji). Remark 1.3. If one imagines that the light gray area in Figure1 corresponds to the percolation cluster at some time, its border (shown in black in Figure1) evolves as the height function in the Totally Asymmetric Simple Exclusion process on the positive integers with open boundary condition, so that all our results could be equivalently phrased in terms of the latter model.
1.1. Previous results in (half-space) LPP. The history of fluctuation results in last passage percolation goes back to the solution to Ulam’s problem about the asymptotic distribution of the size of the longest increasing subsequence in a uniformly random permutation [4]. A proof of the nature of fluctuations in exponential distribution LPP is provided in [29]. On a half-space, [5,6,7] studies the longest increasing subsequence in random involutions. This is equivalent to a half-space LPP problem since for an involution σ ∈ Sn, the graph of i 7→ σ(i) is symmetric with respect to the first diagonal. Actually, [5,6,7] treat both the longest increasing subsequences of a random involution and symmetrized LPP with geometric weights. That work contains the geometric weight half-space LPP analogue of Theorem 1.4. The methods used therein only worked when restricted to the one point distribution of passage times exactly along the diagonal. Further work on half-space LPP was undertaken in [38], where the results are stated there in terms of the discrete polynuclear growth (PNG) model rather than the equivalent half-space LPP with geometric weights. The framework of [38] allows to study the correlation kernel corresponding to multiple points in the space direction, with or without nucleations at the origin. In the large scale limit, the authors performed a non-rigorous critical point derivation of the limiting correlation kernel in certain regimes and found Airy-type process and various limiting crossover kernels near the origin, much in the same spirit as our results (see the discussion about these results in Section 1.3). 2 1.2. Previous methods. A key property in the solution of Ulam’s problem in [4] was the relation to the RSK algorithm which reveals a beautiful algebraic structure that can be generalized using the formalism of Schur measures [33] and Schur processes [34]. RSK is just one of a variety of Markov dynamics on sequences of partitions which preserves the Schur process [17, 13]. Studying these dynamics gives rise to a number of interesting probabilistic models related to Schur processes. In the early works of [4,6] the convergence to the Tracy-Widom distributions was proved by Riemann- Hilbert problem asymptotic analysis. Subsequently Fredholm determinants became the preferred vehicle for asymptotic analysis. In a half-space, [5] applied RSK to symmetric matrices. Subsequently, [37] (see also [26]) showed that geometric weight half-space LPP is a Pfaffian point process (see Section 4.1) and computed the correlation kernel. Later, [21] defined Pfaffian Schur processes generalizing the probability measures defined in [37] and [38]. The Pfaffian Schur process is an analogue of the Schur process for symmetrized problems. 1.3. Our methods and new results. We consider Markov dynamics similar to those of [17] and show that they preserve Pfaffian Schur processes. This enables us to show how half-space LPP with geometric weights is a marginal of the Pfaffian Schur process. We then apply a general formula giving the correlation kernel of Pfaffian Schur processes [21] to study the model’s multipoint distributions. We degenerate our model and formulas to the case of exponential weight half-space LPP (previous works [5,6,7, 38] only considered geometric weights). We perform an asymptotic analysis of the Fredholm Pfaffians characterizing the distribution of passage times in this model, which leads to Theorems 1.4 and 1.5. We also study the k-point distribution of passage times at a distance O(n2/3) away from the diagonal, and introduce a new two-parameter family of crossover distributions when the rate of the weights on the diagonal are simultaneously scaled close to their critical value. This generalizes the crossover distributions found in [25, 38]. The asymptotics of fluctuations away from the diagonal were already studied in [38] for the PNG model on a half-space (equivalently the geometric weight half-space LPP). The asymptotic analysis there was non-rigorous, at the level of studying the behavior of integrals around their critical points without controlling the decay of tails of integrals. The proof of the first part of Theorem 1.4 (in Section5) required a new idea which we believe is the most novel technical contribution of this paper: When α > 1/2, H(n, n) is the maximum of a simple Pfaffian point process which converges as n → ∞ to a point process where all points have multiplicity 2. The limit of the correlation kernel does not exist2 as a function (it does have a formal limit involving the derivative of the Dirac delta function). Nevertheless, a careful reordering and non-trivial cancellation of terms in the Fredholm Pfaffian expansions allows us to study the law of the maximum of this limiting point process, and ultimately find that it corresponds to the GSE Tracy-Widom distribution. We actually provide a general scheme for how this works for random matrix type kernels in which simple Pfaffian point processes limit to doubled ones (see Section 5.1). Theorem 5.3 of [38] derives a formula for certain crossover distributions in half-space geometric LPP. This crossover distribution, introduced in [25], depends on a parameter τ (which is denoted η in our Theorem 1.9) and it corresponds to a natural transition ensemble between GSE when τ → 0 and GUE when τ → +∞. Although the collision of eigenvalues should occur in the GSE limit of this crossover, this point was left unaddressed in previous literature and the convergence of the crossover kernel to the GSE kernel was not proved in [25] (see our Proposition 6.12). The formulas for limiting kernels, given by [25, (4.41), (4.44), (4.47)] or [38, (5.37)-(5.40)], make sense so long as τ > 0. Notice a difference of integration domain between [25, (4.44)] and [38, (5.39)]. Our Theorem 1.9 is the analogue of [38, Theorem 5.3] for half-space exponential LPP and we recover
2For a non-simple point process, the correlation functions – Radon-Nikodym derivatives of the factorial moment measures – generically do not exist (see Definitions in Section 4.1). 3 3 the exact same kernel. When τ = 0 the formula for I4 in [38, (5.39)] involves divergent integrals , though if one uses the formal identity R Ai(x + r)Ai(y + r)dr = δ(x − y) then it is possible to R rewrite the kernel in terms of an expression involving the derivative of the Dirac delta function, and the expression formally matches with the kernel K∞ introduced in Section 5.1. From that (formal) kernel it is non-trivial to match the Fredholm Pfaffian with a known formula for the GSE distribution, this matching does not seem to have been made previously in the literature and we explain it in Section 5.1. In random matrix theory, a similar phenomenon of collisions of eigenvalues occurs in the limit from the discrete symplectic ensemble to the GSE [22]. However, [22] used averages of character- istic polynomials to characterize the point process, and computed the correlation kernels from the characteristic polynomials only after the limit, so that collisions of eigenvalues were not an issue. A random matrix ensemble which crossovers between GSE and GUE was studied in [25].
1.4. Other models related to KPZ growth in a half-space. A positive temperature analogue of LPP – directed random polymers – has also been studied. The log-gamma polymer in a half- space (which converges to half-space LPP with exponential weights in the zero temperature limit) is considered in [36] where an exact formula is derived for the distribution of the partition function in terms of Whittaker functions. Some of these formulas are not yet proved and presently there has not been a successful asymptotic analysis preformed of them. Within physics, [28] and [14] studies the continuum directed random polymer (equivalently the KPZ equation) in a half-space with a repulsive and refecting (respectively) barrier and derives GSE Tracy-Widom limiting statistics. Both works are mathematically non-rigorous due to the ill-posed moment problem and certain unproved forms of the conjectured Bethe ansatz completeness of the half-space delta Bose gas. On the side of particle systems, half-space ASEP has first been studied in [41], but the resulting formulas were not amenable to asymptotic analysis. More recently, [9] derives exact Pfaffian formulas for the half-space stochastic six-vertex model, and proves GOE asymptotics for the current at the origin in half-space ASEP, for a certain boundary condition. The exact formulas involve the correlation kernel of the Pfaffian Schur process and the asymptotic analysis in [9] uses some of the ideas developed here.
1.5. Main results. We now state the limit theorems that constitute the main results of this paper. They involve the GOE, GSE and GUE Tracy-Widom distributions, respectively characterized by the distribution functions FGOE, FGSE and FGUE given in Section2 and the standard Gaussian distribution function denoted by G(x). Theorem 1.4. The last passage time on the diagonal H(n, n) satisfies the following limit theorems, depending on the rate α of the weights on the diagonal. (1) For α > 1/2, H(n, n) − 4n lim P < x = FGSE (x) . n→∞ 24/3n1/3 (2) For α = 1/2, H(n, n) − 4n lim P < x = FGOE (x) . n→∞ 24/3n1/3 (3) For α < 1/2, n ! H(n, n) − α(1−α) lim P < x = G(x), n→∞ σn1/2
3The saddle-point analysis in the proof of Theorem 5.3 in [38] is valid only when τ > 0. Indeed, [38, (5.44)] requires τ1 + τ2 > η1 + η2 and one needs that η1, η2 > 0 as in the proof of [38, Theorem 4.2]. 4 (1−2α)1/2 where σ = α(1−α) .
This extends the results of [5,6] on the model with geometric weights. The first part of Theorem 1.4 is proved in Section5, while the second and third parts are proved in Section6. Section 6.1 includes an explanation for why the transition occurs at α = 1/2, using a property of the Pfaffian Schur measure (Prop. 3.4). Far away from the diagonal, the limit theorem satisfied by H(n, m) coincides with that of the (unsymmetrized) full-quadrant model. √ √κ Theorem 1.5. For any κ ∈ (0, 1) and α > 1+ κ , we have that √ H(n, κn) − (1 + κ)2n lim P < x = FGUE(x), n→∞ σn1/3 √ (1+ κ)4/3 where σ = √ . κ1/3 This extends the results of [38] on the geometric model to the exponential case and to allow any boundary parameter. Theorem 1.5 is proved in Section 6.3. Remark 1.6. Our asymptotic analyses could be extended to show that the fluctuations of H(n, κn) √ √ for κ ∈ (0, 1) follow the BBP transition [3] when α = κ/(1+ κ)+O(n−1/3), and in particular are 2 √ √ distributed according to FGOE when α = κ/(1 + κ). The BBP transition would also occur if one varies the rate of exponential weights in the first rows of the lattice. Regarding H(n, n), it is not clear if the higher order phase transitions would coincide with the spiked GSE [12, 42]. Our results rely on an asymptotic analysis of the following formula for the joint distribution of passage times. It involves the Fredholm Pfaffian (see Section2 for background on Fredholm Pfaf- fians) of a matrix-valued correlation kernel Kexp defined in Section 4.4 where the next proposition is proved.
Proposition 1.7. For any h1, . . . , hk > 0 and integers 0 < n1 < n2 < ··· < nk and m1 > m2 > ··· > mk such that ni > mi for all i, we have that k ! \ exp P H(ni, mi) < hi = Pf J − K 2 , L Dk(h1,...,hk) i=1 where the R.H.S is a Fredholm Pfaffian (Definition 2.3) on the domain
Dk(h1, . . . , hk) = {(i, x) ∈ {1, . . . , k} × R : x > hi}.
In the one point case k = 1, and for n1 = m1, one could alternatively characterize the probability distribution of H(n, n) by taking the exponential limit of the formulas in [5,6] using for instance the results from [1]. We can generalize the results of Theorems 1.4 and 1.5 by allowing the parameters κ and α to vary in regions of size n−1/3 around the critical values α = 1/2, κ = 1. In the limit, we obtain a new two-parametric family of probability distributions. More generally, we will compute the finite dimensional marginals of Airy-like processes in various ranges of α and κ. KPZ scaling predictions (together with Theorem 1.5) suggests to define H n + n2/3ξη, n − n2/3ξη − 4n + n1/3ξ2η2 H (η) = , n σn1/3 5 GSE GUE $ α = +∞ 1+22/3$k−1/3 α = 2 GSE n = k + 22/3k2/3η 2/3 2/3 m = k − 2 k η Lcross(·; $, η) GUE α = 1/2 GOE GUE η GOE2 Gaussian α = 0 n = 1 n = +∞ m m GUE
Gaussian GOE2
Figure 2. Phase diagram of the fluctuations of H(n, m) as n → ∞ when α and the ratio n/m varies. The gray area corresponds to a region of the parameter space where the fluctuations are on the scale n1/2 and Gaussian. The bounding curve (where fluctuations are expected to be Tracy-Widom GOE2 cf Remark 1.6) asymptotes to zero as n/m goes to +∞. The crossover distribution Lcross(·; $, η) is defined in Definition 2.9 and describes the fluctuations in the vicinity of n/m = 1 and α = 1/2.
4/3 2/3 where η > 0, σ = 2 and ξ = 2 . Let us scale α as 1 + 2σ−1$n−1/3 α = 2 where $ ∈ R is a free parameter. The limiting joint distribution of multiple points is characterized by a new crossover kernel Kcross that we introduce in Section 2.5.
1+22/3$n−1/3 Theorem 1.8. For 0 6 η1 < ··· < ηk, $ ∈ R, setting α = 2 , we have that k ! \ cross lim Hn(ηi) < xi = Pf J − K 2 . n→∞ P L (Dk(x1,...,xk)) i=1
The proof of this result is very similar with the proof of Theorem 1.5. We provide the details of the computations in [2]. In the one-point case (k = 1), this yields a new probability distribution Lcross, depending on two parameters $, η, (see Definition 2.9), which realizes a two-dimensional crossover between GSE, GUE and GOE distributions (see Figure2 and details in Section 2.5). When η = 0, we recover the probability distribution F (x; w) from [6, Definition 4]. When $ = 0, the correlation kernel Kcross degenerates to the orthogonal-unitary transition kernel obtained in [25] and Fcross(x; 0, η) realizes a crossover between FGOE for η = 0 and FGUE for η → +∞. An analogous result was obtained in [38, Theorem 4.2] for the half-space PNG model. In the case when α > 1/2 is fixed, the joint distribution of passage-times is governed by the so-called symplectic-unitary transition [25].
Theorem 1.9. For α > 1/2 and 0 < η1 < ··· < ηk, we have that SU lim (Hn(η1) < x1,...,Hn(ηk) < xk) = Pf J − K 2 , n→∞ P L (Dk(x1,...,xk)) 6 where the kernel KSU is defined in Section 2.5. Theorem 1.9 can be seen as a $ → +∞ degeneration of Theorem 1.8 (see Section 2.5). The proof is very similar with that of Theorem 1.8. We provide the details of the computations in [2]. An analogous result was found for the half-space PNG model without nucleations at the origin [38] although the statement of [38, Theorem 5.3] makes rigorous sense only when τ > 0. The correlation kernel corresponding to the symplectic-unitary transition was studied first in [25] 4. This random matrix ensemble is the point process corresponding to the eigenvalues of a Hermitian complex matrix Xη for η ∈ (0, +∞) with density proportional to
−η 2 −Tr (Xη−e X0) exp 1−e−2η , where X0 is a GSE matrix. Various limits of the symplectic-unitary and orthogonal-unitary tran- sition kernels are considered in [25, Section 5]. In particular [25, Section 5.6] considers the limit as η → 0. While the GOE distribution is recovered in the orthogonal-unitary case, the explanations are missing in the symplectic-unitary case (the scaling argument at the end of [25, Section 5.6] is not the reason why one does not recover KGSE). We provide a rigorous proof in Section 6.5. Remark 1.10. One can also study the limiting n-point distribution of √ H(n + n2/3ξη, κn − n2/3ξη) − (1 + κ)2n − n1/3x2 η 7→ , 24/3n1/3 for a fixed κ ∈ (0, 1) and several values of x. In the n → ∞ limit, one would obtain the extended exp exp exp Airy kernel for K12 and 0 for K11 and K22 but we do not pursue that direction. Outline of the paper. In Section2, we provide convenient Fredholm Pfaffian formulas for the Tracy-Widom distributions and their generalizations. In Section3, we define Pfaffian Schur pro- cesses and construct dynamics preserving them, thus making a connection to half-space LPP (Proposition 3.10). In Section4, we apply a general result of Borodin and Rains giving the cor- relation structure of Pfaffian Schur processes, in order to express the k-point distribution along space-like paths in half-space LPP with geometric (Proposition 4.2) and exponential (Proposition 1.7) weights. In Section5, we discussed the issues related to the multiplicity of the limiting point process, and prove the first part of Theorem 1.4. In Section6, we perform all other asymptotic analysis: we prove limit theorems towards the GUE Tracy-Widom distribution (Theorem 1.5), GOE Tracy-Widom distribution (second part of Theorem 1.4), and Gaussian distribution (third part of 1.4).
Acknowledgements. The authors would like to thank Alexei Borodin and Eric Rains for sharing their insights about Pfaffian Schur processes. We are especially grateful to Alexei Borodin for dis- cussions related to the dynamics preserving Pfaffian Schur processes and for drawing our attention to the reference [22]. We also thank Tomohiro Sasamoto and Takashi Imamura for discussions regarding their paper [38]. Part of this work was done during the stay of J. Baik, G. Barraquand and I. Corwin at the KITP and supported by NSF PHY11-25915. J. Baik was supported in part by NSF grant DMS-1361782. G. Barraquand was supported in part by the LPMA UMR CNRS 7599, Universit´eParis-Diderot–Paris 7 and the Packard Foundation through I. Corwin’s Packard Fellowship. I.Corwin was supported in part by the NSF through DMS-1208998, a Clay Research Fellowship, the Poincar´eChair, and a Packard Fellowship for Science and Engineering.
4 The integration domain in [25, (4.44)] should be R<0 instead of R>0. The correct formula was found in [38, (5.39)]. 7 2. Fredholm Pfaffian formulas Let us introduce a convenient notation that we use throughout the paper to specify integration contours in the complex plane. ϕ Definition 2.1. Let Ca be the union of two semi-infinite rays departing a ∈ C with angles ϕ and −ϕ. We assume that the contour is oriented from a + ∞e−iϕ to a + ∞e+iϕ.
2.1. GUE Tracy-Widom distribution. For a kernel K : X × X → R, we define its Fredholm determinant det(I + K)L2(X,µ) as given by the series expansion ∞ Z Z X 1 k ⊗k det(I + K) 2 = 1 + ··· det K(x , x ) dµ (x . . . x ), L (X,µ) k! i j i,j=1 1 k k=1 X X whenever it converges. We will generally omit the measure µ in the notations and write simply 2 L (X) when the uniform or the Lebesgue measure is considered.
Definition 2.2. The GUE Tracy-Widom distribution, denoted LGUE is a probability distribution on R such that if X ∼ LGUE,
P(X 6 x) = FGUE(x) = det(I − KAi)L2(x,+∞) where KAi is the Airy kernel, Z Z ez3/3−zu 1 KAi(u, v) = dw dz 3 . (1) 2π/3 π/3 ew /3−wv z − w C−1 C1 Throughout the paper, all integrals over a contour of the complex plane will take a factor 1/(2iπ), 1 and thus we use the notation dz := 2iπ dz. 2.2. Fredholm Pfaffian. In order to define the GOE and GSE distribution in a form which is convenient for later purposes, we introduce the concept of Fredholm Pfaffian. We refer to Section 4.1 explaining how Fredholm Pfaffians naturally arise in the study of Pfaffian point processes. Definition 2.3 ([37] Section 8). For a 2 × 2-matrix valued skew-symmetric kernel, K11(x, y) K12(x, y) K(x, y) = , x, y ∈ X, K21(x, y) K22(x, y) we define its Fredholm Pfaffian of K by the series expansion ∞ Z Z k X 1 ⊗k Pf J + K 2 = 1 + ··· Pf K(xi, xj) dµ (x1 . . . xk), (2) L (X,µ) k! i,j=1 k=1 X X provided the series converges, and we recall that for an skew-symmetric 2k × 2k matrix A, its Pfaffian is defined by 1 X Pf(A) = sign(σ)a a . . . a . (3) 2kk! σ(1)σ(2) σ(3)σ(4) σ(2k−1)σ(2k) σ∈S2k The operator J is defined by its kernel J, where 0 1 J(x, y) = 1 . x=y −1 0
8 Remark 2.4. A 2 × 2-matrix valued kernel K(x, y) is called skew-symmetric if the 2k × 2k matrix k 2 K(xi, xj) is skew-symmetric. We must consider it as made up of k blocks, each of which i,j=1 has size 2 × 2. Considering 22 blocks of size k × k instead would change the value of its Pfaffian by a factor (−1)k(k−1)/2.
In Sections4,5 and6, we will need to control the convergence of Fredholm Pfaffian series expansions. This can be done using Hadamard’s bound. We recall that for a matrix M of size k, if the (i, j) entry of M is bounded by aibj for all i, j ∈ {1, . . . , k} then
k k/2 Y det[M] 6 k aibi. i=1
Using this inequality and the fact that Pf[A] = pdet[A] for a skew-symmetric matrix A, we have
Lemma 2.5. Let K(x, y) a 2 × 2 matrix valued skew symmetric kernel. Assume that there exist constants C > 0 and constants a > b > 0 such that −ax−ay −ax+by bx+by |K11(x, y)| < Ce , |K12(x, y)| < Ce , |K22(x, y)| < Ce .
Then, for all k ∈ Z>0,
k k Y PfK(x , x ) < (2k)k/2Ck e−(a−b)xi . i j i,j=1 i=1
2.3. GOE Tracy-Widom distribution. The GOE Tracy-Widom distribution, denoted LGOE, is a continuous probability distribution on R. The following is a convenient formula for its cumulative distribution function.
Lemma 2.6. For X ∼ LGOE, GOE FGOE(x) := (X x) = Pf J − K 2 , P 6 L (x,∞) where KGOE is the 2 × 2 matrix valued kernel defined by Z Z GOE z − w z3/3+w3/3−xz−yw K11 (x, y) = dz dw e , π/3 π/3 z + w C1 C1 Z Z GOE GOE w − z z3/3+w3/3−xz−yw K12 (x, y) = −K21 (x, y) = dz dw e , π/3 π/3 2w(z + w) C1 C−1/2 Z Z GOE z − w z3/3+w3/3−xz−yw K22 (x, y) = dz dw e π/3 π/3 4zw(z + w) C1 C1 Z 3 dz Z 3 dz sgn(x − y) + ez /3−zx − ez /3−zy − . π/3 4z π/3 4z 4 C1 C1 Throughout the paper, we adopt the convention that
sgn(x − y) = 1x>y − 1x 9 Proof. According to [40] (see also [23, (2.9) and (6.17)]), the GOE distribution function can be 1 1 5 defined by FGOE = Pf J − K , where K is defined by the matrix kernel Z ∞ Z ∞ 1 0 0 K11(x, y) = dλAi(x + λ)Ai (y + λ) − dλAi(y + λ)Ai (x + λ), 0 0 Z ∞ Z ∞ 1 1 1 K12(x, y) = −K21(y, x) = Ai(x + λ)Ai(y + λ) + Ai(x) dλAi(y − λ) 0 2 0 Z ∞ Z ∞ Z ∞ Z ∞ 1 1 1 K22(x, y) = dλ dµAi(y + µ)Ai(x + λ) − dλ dµAi(x + µ)Ai(y + λ) 4 0 λ 4 0 λ 1 Z +∞ 1 Z +∞ sgn(x − y) − Ai(x + λ)dλ + Ai(y + λ)dλ − . 4 0 4 0 4 Using the contour integral representation of the Airy function Z z3/3−zx Ai(x) = e dz, for any ϕ ∈ (π/6, π/2] and a ∈ >0, (4) ϕ R Ca π/3 0 with integration on C1 for Ai(x + λ) and Ai (x + λ), we readily see that ∞ Z Z Z 3 3 K1 (x, y) = dλ dz dw(z − w)ez /3+w /3−xz−yw−λz−λw 11 π/3 π/3 0 C1 C1 ∞ Z Z 3 3 Z = dz dw(z − w)ez /3+w /3−xz−yw dλe−λz−λw (5) π/3 π/3 C1 C1 0 GOE = K11 (x, y). The exchange of integrations in (5) is justified because z + w has positive real part along the GOE π/3 π/3 contours. Turning to K12 , we first deform the contour C−1/2 for the variable w to C1 . When doing this, we encounter a pole at zero. Thus, taking into account the residue at zero, we find that Z Z GOE 1 z − w z3/3+w3/3−xz−yw K12 (x, y) = Ai(x) + dz dw e 2 π/3 π/3 2w(z + w) C1 C1 π/3 Using again (4) with the contour C1 , we find that Z ∞ Z ∞ GOE 1 1 1 K12 (x, y) = Ai(x) + Ai(x + λ)Ai(y + λ)dλ − Ai(x) dλAi(y + λ) = K12(x, y), 2 0 2 0 R +∞ 1 where in the last equality we have used that 0 Ai(λ)dλ = 1. For K22 we use similarly (4) with π/3 1 GOE integration on C1 for Ai(x + λ) and Ai(x + µ) and get that K22(x, y) = K22 (x, y). 2.4. GSE Tracy-Widom distribution. The GSE Tracy-Widom distribution, denoted LGSE, is a continuous probability distribution on R. Lemma 2.7. For X ∼ LGSE,